Deep learning automatically distinguishes myocarditis patients from normal subjects based on MRI

Fig. 1. Views considered for the detection of myocarditis from CMRI: (a) T2 weighted 2-chamber view, (b) T2 weighted 4-chamber view, (c) LGE 2-chamber view, (d) LGE 4-chamber view, (e) cine bSSFP 2-chamber view, (f) cine bSSFP 3-chamber view, (g) cine bSSFP 4-chamber view, (h) cine bSSFP SAX stack. The middle frame is displayed for multi frame views, and for the multi-slice multi-frame view, we used the middle slice and the middle frame

Fig. 2. Visual illustration using different increments of k for LGE 2-chamber view and LGE 4-chamber view

Fig. 3. Illustration of the multi-frame processing approach using a spatial CNN backbone with six convolutional layers. This diagram depicts the process where the backbone is applied independently to each frame, followed by the concatenation of resulting feature maps from all frames. A subsequent single 2D convolutional layer is then utilized to extract and integrate both spatial and temporal features across the entire frame set

Fig. 4. The five middle slices for the multi slice multi frame view (cine bSSFP SAX stack): (a) slice 5, (b) slice 6, (c) slice 7, (d) slice 8, (e) slice 9. The middle frame is depicted for all 5 slices

Fig. 5. Two sample saliency maps are shown for the LGE 2-chamber view (top row) and the LGE 4-chamber view (bottom row). For each view, the saliency maps are presented with the control subject on the top and the myocarditis case on the bottom

Fig. 6. ROC curvers obtained for the best performing models, i.e., employing FSL on the LGE 2-chamber and LGE 4-chamber views respectively

Fig. 7. Confusion matrix obtained for the model corresponding to the LGE 2-chamber view-based classification.

Fig. 8. Confusion matrix obtained for the model corresponding to the LGE 4-chamber view classification

Fig. 9. Four sample cases for the LGE 2-chamber view: (a) TP, (b) FP, (c) TN, and (d) FN

Fig. 10. Four sample cases for the LGE 2-chamber view: (a) TP, (b) FP, (c) TN, and (d) FN

Fig. 11. ROC curves obtained at the last epoch for the best performing models, i.e., employing FSL on the LGE 2-chamber and LGE 4-chamber views respectively